Carolina Ballroom
Sponsored by Bank of America
Clustering Multiple Outcomes via a Dirichlet Process Prior (303505)
*Amy Allynne LaLonde, University of RochesterTanzy Love, University of Rochester
Environmental exposure effects are often small and difficult to detect given the observational nature of the data. The Seychelles Child Development Study (SCDS) examined the effect of prenatal methylmercury exposure using a battery of tests measuring various aspects of child development. The 20 SCDS test outcomes have been examined in past analyses (Myers, et al., 2003; Thurston et al., 2009; Xiao et al., 2014). Thurston et al., posed a multiple outcomes model with outcomes nested into domains to allow covariate and exposure effects to differ across domains, and outcome-, subject- and domain-specific subject random effects to fully capture correlations among outcomes in the same domain and across different domains. Xiao et al. extended the model by allowing the data to inform partial membership of outcomes into domains. Our model is a middle ground adopting the single domain assignments of outcomes, but using information in the data to cluster the outcomes into domains using a Dirichlet process prior within a Bayesian MCMC (Neal, 2000). Results show more pronounced exposure effects in one of the domains, distinct domain-specific covariate effects, and sensible domain assignments.